sd_models.py

import collections
import os
import sys
import threading

import torch
import re
import safetensors.torch
from omegaconf import OmegaConf, ListConfig
from urllib import request
import ldm.modules.midas as midas

from ldm.util import instantiate_from_config

from modules import paths, shared, modelloader, devices, script_callbacks, sd_vae, sd_disable_initialization, errors, hashes, sd_models_config, sd_unet, sd_models_xl, cache, extra_networks, processing, lowvram, sd_hijack, patches
from modules.timer import Timer
from modules.shared import opts
import tomesd
import numpy as np

model_dir = "Stable-diffusion"
model_path = os.path.abspath(os.path.join(paths.models_path, model_dir))

checkpoints_list = {}
checkpoint_aliases = {}
checkpoint_alisases = checkpoint_aliases  # for compatibility with old name
checkpoints_loaded = collections.OrderedDict()


def replace_key(d, key, new_key, value):
    keys = list(d.keys())

    d[new_key] = value

    if key not in keys:
        return d

    index = keys.index(key)
    keys[index] = new_key

    new_d = {k: d[k] for k in keys}

    d.clear()
    d.update(new_d)
    return d


class CheckpointInfo:
    def __init__(self, filename):
        self.filename = filename
        abspath = os.path.abspath(filename)
        abs_ckpt_dir = os.path.abspath(shared.cmd_opts.ckpt_dir) if shared.cmd_opts.ckpt_dir is not None else None

        self.is_safetensors = os.path.splitext(filename)[1].lower() == ".safetensors"

        if abs_ckpt_dir and abspath.startswith(abs_ckpt_dir):
            name = abspath.replace(abs_ckpt_dir, '')
        elif abspath.startswith(model_path):
            name = abspath.replace(model_path, '')
        else:
            name = os.path.basename(filename)

        if name.startswith("\\") or name.startswith("/"):
            name = name[1:]

        def read_metadata():
            metadata = read_metadata_from_safetensors(filename)
            self.modelspec_thumbnail = metadata.pop('modelspec.thumbnail', None)

            return metadata

        self.metadata = {}
        if self.is_safetensors:
            try:
                self.metadata = cache.cached_data_for_file('safetensors-metadata', "checkpoint/" + name, filename, read_metadata)
            except Exception as e:
                errors.display(e, f"reading metadata for {filename}")

        self.name = name
        self.name_for_extra = os.path.splitext(os.path.basename(filename))[0]
        self.model_name = os.path.splitext(name.replace("/", "_").replace("\\", "_"))[0]
        self.hash = model_hash(filename)

        self.sha256 = hashes.sha256_from_cache(self.filename, f"checkpoint/{name}")
        self.shorthash = self.sha256[0:10] if self.sha256 else None

        self.title = name if self.shorthash is None else f'{name} [{self.shorthash}]'
        self.short_title = self.name_for_extra if self.shorthash is None else f'{self.name_for_extra} [{self.shorthash}]'

        self.ids = [self.hash, self.model_name, self.title, name, self.name_for_extra, f'{name} [{self.hash}]']
        if self.shorthash:
            self.ids += [self.shorthash, self.sha256, f'{self.name} [{self.shorthash}]', f'{self.name_for_extra} [{self.shorthash}]']

    def register(self):
        checkpoints_list[self.title] = self
        for id in self.ids:
            checkpoint_aliases[id] = self

    def calculate_shorthash(self):
        self.sha256 = hashes.sha256(self.filename, f"checkpoint/{self.name}")
        if self.sha256 is None:
            return

        shorthash = self.sha256[0:10]
        if self.shorthash == self.sha256[0:10]:
            return self.shorthash

        self.shorthash = shorthash

        if self.shorthash not in self.ids:
            self.ids += [self.shorthash, self.sha256, f'{self.name} [{self.shorthash}]', f'{self.name_for_extra} [{self.shorthash}]']

        old_title = self.title
        self.title = f'{self.name} [{self.shorthash}]'
        self.short_title = f'{self.name_for_extra} [{self.shorthash}]'

        replace_key(checkpoints_list, old_title, self.title, self)
        self.register()

        return self.shorthash


try:
    # this silences the annoying "Some weights of the model checkpoint were not used when initializing..." message at start.
    from transformers import logging, CLIPModel  # noqa: F401

    logging.set_verbosity_error()
except Exception:
    pass


def setup_model():
    """called once at startup to do various one-time tasks related to SD models"""

    os.makedirs(model_path, exist_ok=True)

    enable_midas_autodownload()
    patch_given_betas()


def checkpoint_tiles(use_short=False):
    return [x.short_title if use_short else x.title for x in checkpoints_list.values()]


def list_models():
    checkpoints_list.clear()
    checkpoint_aliases.clear()

    cmd_ckpt = shared.cmd_opts.ckpt
    if shared.cmd_opts.no_download_sd_model or cmd_ckpt != shared.sd_model_file or os.path.exists(cmd_ckpt):
        model_url = None
    else:
        model_url = f"{shared.hf_endpoint}/runwayml/stable-diffusion-v1-5/resolve/main/v1-5-pruned-emaonly.safetensors"

    model_list = modelloader.load_models(model_path=model_path, model_url=model_url, command_path=shared.cmd_opts.ckpt_dir, ext_filter=[".ckpt", ".safetensors"], download_name="v1-5-pruned-emaonly.safetensors", ext_blacklist=[".vae.ckpt", ".vae.safetensors"])

    if os.path.exists(cmd_ckpt):
        checkpoint_info = CheckpointInfo(cmd_ckpt)
        checkpoint_info.register()

        shared.opts.data['sd_model_checkpoint'] = checkpoint_info.title
    elif cmd_ckpt is not None and cmd_ckpt != shared.default_sd_model_file:
        print(f"Checkpoint in --ckpt argument not found (Possible it was moved to {model_path}: {cmd_ckpt}", file=sys.stderr)

    for filename in model_list:
        checkpoint_info = CheckpointInfo(filename)
        checkpoint_info.register()


re_strip_checksum = re.compile(r"\s*\[[^]]+]\s*$")


def get_closet_checkpoint_match(search_string):
    if not search_string:
        return None

    checkpoint_info = checkpoint_aliases.get(search_string, None)
    if checkpoint_info is not None:
        return checkpoint_info

    found = sorted([info for info in checkpoints_list.values() if search_string in info.title], key=lambda x: len(x.title))
    if found:
        return found[0]

    search_string_without_checksum = re.sub(re_strip_checksum, '', search_string)
    found = sorted([info for info in checkpoints_list.values() if search_string_without_checksum in info.title], key=lambda x: len(x.title))
    if found:
        return found[0]

    return None


def model_hash(filename):
    """old hash that only looks at a small part of the file and is prone to collisions"""

    try:
        with open(filename, "rb") as file:
            import hashlib
            m = hashlib.sha256()

            file.seek(0x100000)
            m.update(file.read(0x10000))
            return m.hexdigest()[0:8]
    except FileNotFoundError:
        return 'NOFILE'


def select_checkpoint():
    """Raises `FileNotFoundError` if no checkpoints are found."""
    model_checkpoint = shared.opts.sd_model_checkpoint

    checkpoint_info = checkpoint_aliases.get(model_checkpoint, None)
    if checkpoint_info is not None:
        return checkpoint_info

    if len(checkpoints_list) == 0:
        error_message = "No checkpoints found. When searching for checkpoints, looked at:"
        if shared.cmd_opts.ckpt is not None:
            error_message += f"\n - file {os.path.abspath(shared.cmd_opts.ckpt)}"
        error_message += f"\n - directory {model_path}"
        if shared.cmd_opts.ckpt_dir is not None:
            error_message += f"\n - directory {os.path.abspath(shared.cmd_opts.ckpt_dir)}"
        error_message += "Can't run without a checkpoint. Find and place a .ckpt or .safetensors file into any of those locations."
        raise FileNotFoundError(error_message)

    checkpoint_info = next(iter(checkpoints_list.values()))
    if model_checkpoint is not None:
        print(f"Checkpoint {model_checkpoint} not found; loading fallback {checkpoint_info.title}", file=sys.stderr)

    return checkpoint_info


checkpoint_dict_replacements_sd1 = {
    'cond_stage_model.transformer.embeddings.': 'cond_stage_model.transformer.text_model.embeddings.',
    'cond_stage_model.transformer.encoder.': 'cond_stage_model.transformer.text_model.encoder.',
    'cond_stage_model.transformer.final_layer_norm.': 'cond_stage_model.transformer.text_model.final_layer_norm.',
}

checkpoint_dict_replacements_sd2_turbo = { # Converts SD 2.1 Turbo from SGM to LDM format.
    'conditioner.embedders.0.': 'cond_stage_model.',
}


def transform_checkpoint_dict_key(k, replacements):
    for text, replacement in replacements.items():
        if k.startswith(text):
            k = replacement + k[len(text):]

    return k


def get_state_dict_from_checkpoint(pl_sd):
    pl_sd = pl_sd.pop("state_dict", pl_sd)
    pl_sd.pop("state_dict", None)

    is_sd2_turbo = 'conditioner.embedders.0.model.ln_final.weight' in pl_sd and pl_sd['conditioner.embedders.0.model.ln_final.weight'].size()[0] == 1024

    sd = {}
    for k, v in pl_sd.items():
        if is_sd2_turbo:
            new_key = transform_checkpoint_dict_key(k, checkpoint_dict_replacements_sd2_turbo)
        else:
            new_key = transform_checkpoint_dict_key(k, checkpoint_dict_replacements_sd1)

        if new_key is not None:
            sd[new_key] = v

    pl_sd.clear()
    pl_sd.update(sd)

    return pl_sd


def read_metadata_from_safetensors(filename):
    import json

    with open(filename, mode="rb") as file:
        metadata_len = file.read(8)
        metadata_len = int.from_bytes(metadata_len, "little")
        json_start = file.read(2)

        assert metadata_len > 2 and json_start in (b'{"', b"{'"), f"{filename} is not a safetensors file"
        json_data = json_start + file.read(metadata_len-2)
        json_obj = json.loads(json_data)

        res = {}
        for k, v in json_obj.get("__metadata__", {}).items():
            res[k] = v
            if isinstance(v, str) and v[0:1] == '{':
                try:
                    res[k] = json.loads(v)
                except Exception:
                    pass

        return res


def read_state_dict(checkpoint_file, print_global_state=False, map_location=None):
    _, extension = os.path.splitext(checkpoint_file)
    if extension.lower() == ".safetensors":
        device = map_location or shared.weight_load_location or devices.get_optimal_device_name()

        if not shared.opts.disable_mmap_load_safetensors:
            pl_sd = safetensors.torch.load_file(checkpoint_file, device=device)
        else:
            pl_sd = safetensors.torch.load(open(checkpoint_file, 'rb').read())
            pl_sd = {k: v.to(device) for k, v in pl_sd.items()}
    else:
        pl_sd = torch.load(checkpoint_file, map_location=map_location or shared.weight_load_location)

    if print_global_state and "global_step" in pl_sd:
        print(f"Global Step: {pl_sd['global_step']}")

    sd = get_state_dict_from_checkpoint(pl_sd)
    return sd


def get_checkpoint_state_dict(checkpoint_info: CheckpointInfo, timer):
    sd_model_hash = checkpoint_info.calculate_shorthash()
    timer.record("calculate hash")

    if checkpoint_info in checkpoints_loaded:
        # use checkpoint cache
        print(f"Loading weights [{sd_model_hash}] from cache")
        # move to end as latest
        checkpoints_loaded.move_to_end(checkpoint_info)
        return checkpoints_loaded[checkpoint_info]

    print(f"Loading weights [{sd_model_hash}] from {checkpoint_info.filename}")
    res = read_state_dict(checkpoint_info.filename)
    timer.record("load weights from disk")

    return res


class SkipWritingToConfig:
    """This context manager prevents load_model_weights from writing checkpoint name to the config when it loads weight."""

    skip = False
    previous = None

    def __enter__(self):
        self.previous = SkipWritingToConfig.skip
        SkipWritingToConfig.skip = True
        return self

    def __exit__(self, exc_type, exc_value, exc_traceback):
        SkipWritingToConfig.skip = self.previous


def check_fp8(model):
    if model is None:
        return None
    if devices.get_optimal_device_name() == "mps":
        enable_fp8 = False
    elif shared.opts.fp8_storage == "Enable":
        enable_fp8 = True
    elif getattr(model, "is_sdxl", False) and shared.opts.fp8_storage == "Enable for SDXL":
        enable_fp8 = True
    else:
        enable_fp8 = False
    return enable_fp8


def load_model_weights(model, checkpoint_info: CheckpointInfo, state_dict, timer):
    sd_model_hash = checkpoint_info.calculate_shorthash()
    timer.record("calculate hash")

    if devices.fp8:
        # prevent model to load state dict in fp8
        model.half()

    if not SkipWritingToConfig.skip:
        shared.opts.data["sd_model_checkpoint"] = checkpoint_info.title

    if state_dict is None:
        state_dict = get_checkpoint_state_dict(checkpoint_info, timer)

    model.is_sdxl = hasattr(model, 'conditioner')
    model.is_sd2 = not model.is_sdxl and hasattr(model.cond_stage_model, 'model')
    model.is_sd1 = not model.is_sdxl and not model.is_sd2
    model.is_ssd = model.is_sdxl and 'model.diffusion_model.middle_block.1.transformer_blocks.0.attn1.to_q.weight' not in state_dict.keys()
    if model.is_sdxl:
        sd_models_xl.extend_sdxl(model)

    if model.is_ssd:
        sd_hijack.model_hijack.convert_sdxl_to_ssd(model)

    if shared.opts.sd_checkpoint_cache > 0:
        # cache newly loaded model
        checkpoints_loaded[checkpoint_info] = state_dict.copy()

    model.load_state_dict(state_dict, strict=False)
    timer.record("apply weights to model")

    del state_dict

    if shared.cmd_opts.opt_channelslast:
        model.to(memory_format=torch.channels_last)
        timer.record("apply channels_last")

    if shared.cmd_opts.no_half:
        model.float()
        model.alphas_cumprod_original = model.alphas_cumprod
        devices.dtype_unet = torch.float32
        timer.record("apply float()")
    else:
        vae = model.first_stage_model
        depth_model = getattr(model, 'depth_model', None)

        # with --no-half-vae, remove VAE from model when doing half() to prevent its weights from being converted to float16
        if shared.cmd_opts.no_half_vae:
            model.first_stage_model = None
        # with --upcast-sampling, don't convert the depth model weights to float16
        if shared.cmd_opts.upcast_sampling and depth_model:
            model.depth_model = None

        alphas_cumprod = model.alphas_cumprod
        model.alphas_cumprod = None
        model.half()
        model.alphas_cumprod = alphas_cumprod
        model.alphas_cumprod_original = alphas_cumprod
        model.first_stage_model = vae
        if depth_model:
            model.depth_model = depth_model

        devices.dtype_unet = torch.float16
        timer.record("apply half()")

    apply_alpha_schedule_override(model)

    for module in model.modules():
        if hasattr(module, 'fp16_weight'):
            del module.fp16_weight
        if hasattr(module, 'fp16_bias'):
            del module.fp16_bias

    if check_fp8(model):
        devices.fp8 = True
        first_stage = model.first_stage_model
        model.first_stage_model = None
        for module in model.modules():
            if isinstance(module, (torch.nn.Conv2d, torch.nn.Linear)):
                if shared.opts.cache_fp16_weight:
                    module.fp16_weight = module.weight.data.clone().cpu().half()
                    if module.bias is not None:
                        module.fp16_bias = module.bias.data.clone().cpu().half()
                module.to(torch.float8_e4m3fn)
        model.first_stage_model = first_stage
        timer.record("apply fp8")
    else:
        devices.fp8 = False

    devices.unet_needs_upcast = shared.cmd_opts.upcast_sampling and devices.dtype == torch.float16 and devices.dtype_unet == torch.float16

    model.first_stage_model.to(devices.dtype_vae)
    timer.record("apply dtype to VAE")

    # clean up cache if limit is reached
    while len(checkpoints_loaded) > shared.opts.sd_checkpoint_cache:
        checkpoints_loaded.popitem(last=False)

    model.sd_model_hash = sd_model_hash
    model.sd_model_checkpoint = checkpoint_info.filename
    model.sd_checkpoint_info = checkpoint_info
    shared.opts.data["sd_checkpoint_hash"] = checkpoint_info.sha256

    if hasattr(model, 'logvar'):
        model.logvar = model.logvar.to(devices.device)  # fix for training

    sd_vae.delete_base_vae()
    sd_vae.clear_loaded_vae()
    vae_file, vae_source = sd_vae.resolve_vae(checkpoint_info.filename).tuple()
    sd_vae.load_vae(model, vae_file, vae_source)
    timer.record("load VAE")


def enable_midas_autodownload():
    """
    Gives the ldm.modules.midas.api.load_model function automatic downloading.

    When the 512-depth-ema model, and other future models like it, is loaded,
    it calls midas.api.load_model to load the associated midas depth model.
    This function applies a wrapper to download the model to the correct
    location automatically.
    """

    midas_path = os.path.join(paths.models_path, 'midas')

    # stable-diffusion-stability-ai hard-codes the midas model path to
    # a location that differs from where other scripts using this model look.
    # HACK: Overriding the path here.
    for k, v in midas.api.ISL_PATHS.items():
        file_name = os.path.basename(v)
        midas.api.ISL_PATHS[k] = os.path.join(midas_path, file_name)

    midas_urls = {
        "dpt_large": "https://github.com/intel-isl/DPT/releases/download/1_0/dpt_large-midas-2f21e586.pt",
        "dpt_hybrid": "https://github.com/intel-isl/DPT/releases/download/1_0/dpt_hybrid-midas-501f0c75.pt",
        "midas_v21": "https://github.com/AlexeyAB/MiDaS/releases/download/midas_dpt/midas_v21-f6b98070.pt",
        "midas_v21_small": "https://github.com/AlexeyAB/MiDaS/releases/download/midas_dpt/midas_v21_small-70d6b9c8.pt",
    }

    midas.api.load_model_inner = midas.api.load_model

    def load_model_wrapper(model_type):
        path = midas.api.ISL_PATHS[model_type]
        if not os.path.exists(path):
            if not os.path.exists(midas_path):
                os.mkdir(midas_path)

            print(f"Downloading midas model weights for {model_type} to {path}")
            request.urlretrieve(midas_urls[model_type], path)
            print(f"{model_type} downloaded")

        return midas.api.load_model_inner(model_type)

    midas.api.load_model = load_model_wrapper


def patch_given_betas():
    import ldm.models.diffusion.ddpm

    def patched_register_schedule(*args, **kwargs):
        """a modified version of register_schedule function that converts plain list from Omegaconf into numpy"""

        if isinstance(args[1], ListConfig):
            args = (args[0], np.array(args[1]), *args[2:])

        original_register_schedule(*args, **kwargs)

    original_register_schedule = patches.patch(__name__, ldm.models.diffusion.ddpm.DDPM, 'register_schedule', patched_register_schedule)


def repair_config(sd_config):

    if not hasattr(sd_config.model.params, "use_ema"):
        sd_config.model.params.use_ema = False

    if hasattr(sd_config.model.params, 'unet_config'):
        if shared.cmd_opts.no_half:
            sd_config.model.params.unet_config.params.use_fp16 = False
        elif shared.cmd_opts.upcast_sampling:
            sd_config.model.params.unet_config.params.use_fp16 = True

    if getattr(sd_config.model.params.first_stage_config.params.ddconfig, "attn_type", None) == "vanilla-xformers" and not shared.xformers_available:
        sd_config.model.params.first_stage_config.params.ddconfig.attn_type = "vanilla"

    # For UnCLIP-L, override the hardcoded karlo directory
    if hasattr(sd_config.model.params, "noise_aug_config") and hasattr(sd_config.model.params.noise_aug_config.params, "clip_stats_path"):
        karlo_path = os.path.join(paths.models_path, 'karlo')
        sd_config.model.params.noise_aug_config.params.clip_stats_path = sd_config.model.params.noise_aug_config.params.clip_stats_path.replace("checkpoints/karlo_models", karlo_path)

    # Do not use checkpoint for inference.
    # This helps prevent extra performance overhead on checking parameters.
    # The perf overhead is about 100ms/it on 4090 for SDXL.
    if hasattr(sd_config.model.params, "network_config"):
        sd_config.model.params.network_config.params.use_checkpoint = False
    if hasattr(sd_config.model.params, "unet_config"):
        sd_config.model.params.unet_config.params.use_checkpoint = False


def rescale_zero_terminal_snr_abar(alphas_cumprod):
    alphas_bar_sqrt = alphas_cumprod.sqrt()

    # Store old values.
    alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone()
    alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone()

    # Shift so the last timestep is zero.
    alphas_bar_sqrt -= (alphas_bar_sqrt_T)

    # Scale so the first timestep is back to the old value.
    alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T)

    # Convert alphas_bar_sqrt to betas
    alphas_bar = alphas_bar_sqrt ** 2  # Revert sqrt
    alphas_bar[-1] = 4.8973451890853435e-08
    return alphas_bar


def apply_alpha_schedule_override(sd_model, p=None):
    """
    Applies an override to the alpha schedule of the model according to settings.
    - downcasts the alpha schedule to half precision
    - rescales the alpha schedule to have zero terminal SNR
    """

    if not hasattr(sd_model, 'alphas_cumprod') or not hasattr(sd_model, 'alphas_cumprod_original'):
        return

    sd_model.alphas_cumprod = sd_model.alphas_cumprod_original.to(shared.device)

    if opts.use_downcasted_alpha_bar:
        if p is not None:
            p.extra_generation_params['Downcast alphas_cumprod'] = opts.use_downcasted_alpha_bar
        sd_model.alphas_cumprod = sd_model.alphas_cumprod.half().to(shared.device)

    if opts.sd_noise_schedule == "Zero Terminal SNR":
        if p is not None:
            p.extra_generation_params['Noise Schedule'] = opts.sd_noise_schedule
        sd_model.alphas_cumprod = rescale_zero_terminal_snr_abar(sd_model.alphas_cumprod).to(shared.device)


sd1_clip_weight = 'cond_stage_model.transformer.text_model.embeddings.token_embedding.weight'
sd2_clip_weight = 'cond_stage_model.model.transformer.resblocks.0.attn.in_proj_weight'
sdxl_clip_weight = 'conditioner.embedders.1.model.ln_final.weight'
sdxl_refiner_clip_weight = 'conditioner.embedders.0.model.ln_final.weight'


class SdModelData:
    def __init__(self):
        self.sd_model = None
        self.loaded_sd_models = []
        self.was_loaded_at_least_once = False
        self.lock = threading.Lock()

    def get_sd_model(self):
        if self.was_loaded_at_least_once:
            return self.sd_model

        if self.sd_model is None:
            with self.lock:
                if self.sd_model is not None or self.was_loaded_at_least_once:
                    return self.sd_model

                try:
                    load_model()

                except Exception as e:
                    errors.display(e, "loading stable diffusion model", full_traceback=True)
                    print("", file=sys.stderr)
                    print("Stable diffusion model failed to load", file=sys.stderr)
                    self.sd_model = None

        return self.sd_model

    def set_sd_model(self, v, already_loaded=False):
        self.sd_model = v
        if already_loaded:
            sd_vae.base_vae = getattr(v, "base_vae", None)
            sd_vae.loaded_vae_file = getattr(v, "loaded_vae_file", None)
            sd_vae.checkpoint_info = v.sd_checkpoint_info

        try:
            self.loaded_sd_models.remove(v)
        except ValueError:
            pass

        if v is not None:
            self.loaded_sd_models.insert(0, v)


model_data = SdModelData()


def get_empty_cond(sd_model):

    p = processing.StableDiffusionProcessingTxt2Img()
    extra_networks.activate(p, {})

    if hasattr(sd_model, 'conditioner'):
        d = sd_model.get_learned_conditioning([""])
        return d['crossattn']
    else:
        return sd_model.cond_stage_model([""])


def send_model_to_cpu(m):
    if m.lowvram:
        lowvram.send_everything_to_cpu()
    else:
        m.to(devices.cpu)

    devices.torch_gc()


def model_target_device(m):
    if lowvram.is_needed(m):
        return devices.cpu
    else:
        return devices.device


def send_model_to_device(m):
    lowvram.apply(m)

    if not m.lowvram:
        m.to(shared.device)


def send_model_to_trash(m):
    m.to(device="meta")
    devices.torch_gc()


def load_model(checkpoint_info=None, already_loaded_state_dict=None):
    from modules import sd_hijack
    checkpoint_info = checkpoint_info or select_checkpoint()

    timer = Timer()

    if model_data.sd_model:
        send_model_to_trash(model_data.sd_model)
        model_data.sd_model = None
        devices.torch_gc()

    timer.record("unload existing model")

    if already_loaded_state_dict is not None:
        state_dict = already_loaded_state_dict
    else:
        state_dict = get_checkpoint_state_dict(checkpoint_info, timer)

    checkpoint_config = sd_models_config.find_checkpoint_config(state_dict, checkpoint_info)
    clip_is_included_into_sd = any(x for x in [sd1_clip_weight, sd2_clip_weight, sdxl_clip_weight, sdxl_refiner_clip_weight] if x in state_dict)

    timer.record("find config")

    sd_config = OmegaConf.load(checkpoint_config)
    repair_config(sd_config)

    timer.record("load config")

    print(f"Creating model from config: {checkpoint_config}")

    sd_model = None
    try:
        with sd_disable_initialization.DisableInitialization(disable_clip=clip_is_included_into_sd or shared.cmd_opts.do_not_download_clip):
            with sd_disable_initialization.InitializeOnMeta():
                sd_model = instantiate_from_config(sd_config.model)

    except Exception as e:
        errors.display(e, "creating model quickly", full_traceback=True)

    if sd_model is None:
        print('Failed to create model quickly; will retry using slow method.', file=sys.stderr)

        with sd_disable_initialization.InitializeOnMeta():
            sd_model = instantiate_from_config(sd_config.model)

    sd_model.used_config = checkpoint_config

    timer.record("create model")

    if shared.cmd_opts.no_half:
        weight_dtype_conversion = None
    else:
        weight_dtype_conversion = {
            'first_stage_model': None,
            'alphas_cumprod': None,
            '': torch.float16,
        }

    with sd_disable_initialization.LoadStateDictOnMeta(state_dict, device=model_target_device(sd_model), weight_dtype_conversion=weight_dtype_conversion):
        load_model_weights(sd_model, checkpoint_info, state_dict, timer)
    timer.record("load weights from state dict")

    send_model_to_device(sd_model)
    timer.record("move model to device")

    sd_hijack.model_hijack.hijack(sd_model)

    timer.record("hijack")

    sd_model.eval()
    model_data.set_sd_model(sd_model)
    model_data.was_loaded_at_least_once = True

    sd_hijack.model_hijack.embedding_db.load_textual_inversion_embeddings(force_reload=True)  # Reload embeddings after model load as they may or may not fit the model

    timer.record("load textual inversion embeddings")

    script_callbacks.model_loaded_callback(sd_model)

    timer.record("scripts callbacks")

    with devices.autocast(), torch.no_grad():
        sd_model.cond_stage_model_empty_prompt = get_empty_cond(sd_model)

    timer.record("calculate empty prompt")

    print(f"Model loaded in {timer.summary()}.")

    return sd_model


def reuse_model_from_already_loaded(sd_model, checkpoint_info, timer):
    """
    Checks if the desired checkpoint from checkpoint_info is not already loaded in model_data.loaded_sd_models.
    If it is loaded, returns that (moving it to GPU if necessary, and moving the currently loadded model to CPU if necessary).
    If not, returns the model that can be used to load weights from checkpoint_info's file.
    If no such model exists, returns None.
    Additionally deletes loaded models that are over the limit set in settings (sd_checkpoints_limit).
    """

    if sd_model is not None and sd_model.sd_checkpoint_info.filename == checkpoint_info.filename:
        return sd_model

    if shared.opts.sd_checkpoints_keep_in_cpu:
        send_model_to_cpu(sd_model)
        timer.record("send model to cpu")

    already_loaded = None
    for i in reversed(range(len(model_data.loaded_sd_models))):
        loaded_model = model_data.loaded_sd_models[i]
        if loaded_model.sd_checkpoint_info.filename == checkpoint_info.filename:
            already_loaded = loaded_model
            continue

        if len(model_data.loaded_sd_models) > shared.opts.sd_checkpoints_limit > 0:
            print(f"Unloading model {len(model_data.loaded_sd_models)} over the limit of {shared.opts.sd_checkpoints_limit}: {loaded_model.sd_checkpoint_info.title}")
            del model_data.loaded_sd_models[i]
            send_model_to_trash(loaded_model)
            timer.record("send model to trash")

    if already_loaded is not None:
        send_model_to_device(already_loaded)
        timer.record("send model to device")

        model_data.set_sd_model(already_loaded, already_loaded=True)

        if not SkipWritingToConfig.skip:
            shared.opts.data["sd_model_checkpoint"] = already_loaded.sd_checkpoint_info.title
            shared.opts.data["sd_checkpoint_hash"] = already_loaded.sd_checkpoint_info.sha256

        print(f"Using already loaded model {already_loaded.sd_checkpoint_info.title}: done in {timer.summary()}")
        sd_vae.reload_vae_weights(already_loaded)
        return model_data.sd_model
    elif shared.opts.sd_checkpoints_limit > 1 and len(model_data.loaded_sd_models) < shared.opts.sd_checkpoints_limit:
        print(f"Loading model {checkpoint_info.title} ({len(model_data.loaded_sd_models) + 1} out of {shared.opts.sd_checkpoints_limit})")

        model_data.sd_model = None
        load_model(checkpoint_info)
        return model_data.sd_model
    elif len(model_data.loaded_sd_models) > 0:
        sd_model = model_data.loaded_sd_models.pop()
        model_data.sd_model = sd_model

        sd_vae.base_vae = getattr(sd_model, "base_vae", None)
        sd_vae.loaded_vae_file = getattr(sd_model, "loaded_vae_file", None)
        sd_vae.checkpoint_info = sd_model.sd_checkpoint_info

        print(f"Reusing loaded model {sd_model.sd_checkpoint_info.title} to load {checkpoint_info.title}")
        return sd_model
    else:
        return None


def reload_model_weights(sd_model=None, info=None, forced_reload=False):
    checkpoint_info = info or select_checkpoint()

    timer = Timer()

    if not sd_model:
        sd_model = model_data.sd_model

    if sd_model is None:  # previous model load failed
        current_checkpoint_info = None
    else:
        current_checkpoint_info = sd_model.sd_checkpoint_info
        if check_fp8(sd_model) != devices.fp8:
            # load from state dict again to prevent extra numerical errors
            forced_reload = True
        elif sd_model.sd_model_checkpoint == checkpoint_info.filename and not forced_reload:
            return sd_model

    sd_model = reuse_model_from_already_loaded(sd_model, checkpoint_info, timer)
    if not forced_reload and sd_model is not None and sd_model.sd_checkpoint_info.filename == checkpoint_info.filename:
        return sd_model

    if sd_model is not None:
        sd_unet.apply_unet("None")
        send_model_to_cpu(sd_model)
        sd_hijack.model_hijack.undo_hijack(sd_model)

    state_dict = get_checkpoint_state_dict(checkpoint_info, timer)

    checkpoint_config = sd_models_config.find_checkpoint_config(state_dict, checkpoint_info)

    timer.record("find config")

    if sd_model is None or checkpoint_config != sd_model.used_config:
        if sd_model is not None:
            send_model_to_trash(sd_model)

        load_model(checkpoint_info, already_loaded_state_dict=state_dict)
        return model_data.sd_model

    try:
        load_model_weights(sd_model, checkpoint_info, state_dict, timer)
    except Exception:
        print("Failed to load checkpoint, restoring previous")
        load_model_weights(sd_model, current_checkpoint_info, None, timer)
        raise
    finally:
        sd_hijack.model_hijack.hijack(sd_model)
        timer.record("hijack")

        if not sd_model.lowvram:
            sd_model.to(devices.device)
            timer.record("move model to device")

        script_callbacks.model_loaded_callback(sd_model)
        timer.record("script callbacks")

    print(f"Weights loaded in {timer.summary()}.")

    model_data.set_sd_model(sd_model)
    sd_unet.apply_unet()

    return sd_model


def unload_model_weights(sd_model=None, info=None):
    send_model_to_cpu(sd_model or shared.sd_model)

    return sd_model


def apply_token_merging(sd_model, token_merging_ratio):
    """
    Applies speed and memory optimizations from tomesd.
    """

    current_token_merging_ratio = getattr(sd_model, 'applied_token_merged_ratio', 0)

    if current_token_merging_ratio == token_merging_ratio:
        return

    if current_token_merging_ratio > 0:
        tomesd.remove_patch(sd_model)

    if token_merging_ratio > 0:
        tomesd.apply_patch(
            sd_model,
            ratio=token_merging_ratio,
            use_rand=False,  # can cause issues with some samplers
            merge_attn=True,
            merge_crossattn=False,
            merge_mlp=False
        )

    sd_model.applied_token_merged_ratio = token_merging_ratio